基于LSTM-Encoder的区域对流层延迟预测模型

Regional tropospheric delay prediction model based on long short-term memory encoder

关键词:  深度学习  长短期记忆编码器  对流层延迟  预测模型

Key words:  Deep learning  Long short-term memory encoder  Tropospheric delay  Prediction model

摘要:

天顶对流层延迟（ZTD）的精确建模对于全球卫星导航系统（GNSS）的实时高精度定位增强至关重要。由于不同地区的大气水汽存在短时变化特性，经验对流层延迟模型在不同地区往往有显著的精度差异，无法满足精确的区域ZTD预测需求。深度学习方法擅长从时间序列数据中学习复杂的非线性模式和依赖关系。利用2023年澳大利亚地区178个连续运行参考站（CORS）的ZTD数据作为真实值，使用长短期记忆编码器（LSTM-Encoder）网络对2023年的第三代全球气温气压模型（GPT3）数据进行建模，并与GPT3模型、欧洲中期天气预报中心（ECMWF）第五代大气再分析数据集（ERA5）模型、人工神经网络（ANN）模型、广义回归神经网络（GRNN）模型和LSTM模型的实验结果进行了比较。结果表明，LSTM-Encoder模型平均偏差接近于0，均方根误差和平均绝对误差分别为14.4 mm和12.4 mm，优于GPT3，ERA5，GRNN，ANN和LSTM模型，均方根误差分别提高了62.2%，12.3%，59.9%，61.0%和60.0%。此外，比较了LSTM-Encoder模型与GPT3和ERA5模型的空间和时间特性，并讨论了不同神经网络方法在不同预报时长下的性能。所提出的预测模型未来可以用于实时精密单点定位（PPP）中ZTD的初始值确定，在观测方程中引入预测的ZTD作为虚拟观测值，促进ZTD与其他待估参数的分离，从而为高精度GNSS定位服务提供理论支持。

Abstract:

Accurate modeling of the zenith tropospheric delay (ZTD) in the troposphere is essential for high-precision real-time positioning in global navigation satellite system (GNSS). Due to the short-term variability characteristics of atmospheric water vapor in different regions, empirical models of tropospheric delay based on meteorological data reanalysis information often show significant differences in accuracy in different regions and cannot meet the demand for accurate regional ZTD predictions. Deep learning methods excel at learning complex nonlinear patterns and dependencies from time series data. In this study, ZTD data from 178 continuously operating reference stations (CORS) in Australia in 2023 are used as the true values. The long short-term memory encoder (LSTM-Encoder) network is applied to model data in 2023 from the third-generation global pressure temperature (GPT3) model. The experimental results of the LSTM-Encoder model are compared with those of the GPT3 model, the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation atmospheric reanalysis dataset (ERA5), the artificial neural network (ANN) model, the generalized regression neural network (GRNN) model and the LSTM model. The results show that the LSTM-Encoder model achieves an average bias close to zero, with a root mean square error (RMSE) of 14.4 mm and a mean absolute error of 12.4 mm, outperforming the GPT3, ERA5, ANN, GRNN and LSTM models. The RMSE is improved by 62.2%, 12.3%, 59.9%, 61.0% and 60.0%, respectively. In addition, the spatial and temporal characteristics of the LSTM-Encoder model are compared with those of the GPT3 and ERA5 models, and the performance of different neural network methods at different prediction durations is discussed. The proposed prediction model can be applied to determine the initial values of ZTD in real-time precise point positioning (PPP) in the future. By introducing the predicted ZTD as virtual observations in the observation equations, the separation of the ZTD from other estimable parameters can be facilitated, thus providing theoretical support for high-precision GNSS positioning services.